基于DEGAN和SHAP的水电机组异常检测研究OA

In hydropower station monitoring systems,fixed threshold methods are commonly used for over-limit alarms,but they exhibit low sensitivity in complex conditions,making early warnings difficult.This paper proposes a feature-enhanced anomaly detection(FEAD-DEGAN)model based on generative adversarial network discriminator and density estimation(DEGAN).Convolution and global average pooling methods optimize the discriminator struc-ture,enhancing time-series feature extraction.The dynamic threshold strategy and kernel density estimation improve detection sensitivity.The model is validated with abnormal oil head swing amplitude data from an axial-flow pump-turbine unit.Compared with Isolation Forest and Autoencoder,the proposed approach shows better performance in anomaly detection success rate and false alarm rate.SHAP quantifies the contribution of monitoring indicators to anomalies,identifying key factors that influence abnormal behavior and enhancing process interpretability.This sup-ports root cause analysis and facilitates the optimization of maintenance strategies,thereby contributing to more effec-tive fault diagnosis and intelligent maintenance.